In this paper, a novel two-time scale task offloading/delivery framework is presented which dynamically manages radio, storage, computing, and cost aspects through multiple Mobile Edge Computing (MEC) servers distributed across Road Side Units (RSUs), Base Stations (BSs), vehicles, and pedestrians. To enhance flexibility, we introduce a smart dynamic task partitioning and cooperative offloading/delivery mechanism, enabling users to offload task partitions to various MEC servers. Uncertainties stemming from users' high mobility and traffic variations may lead to unpredictable task offloading demands and Imperfect Channel State Information (I-CSI), potentially causing unsuccessful task offloading/delivery. In response, we propose a robust and cooperative two-time-scale task offloading/delivery approach, adept at handling demand uncertainty and I-CSI. The problem formulation involves a two-time scale Markov decision process, aimed at minimizing task offloading/delivery costs while maximizing task completion rates. Large time-scale slots incorporate intelligent RSU activation/deactivation, while short time-scale slots focus on task offloading/delivery. To efficiently implement this framework, we propose Federated Reinforcement Learning (FRL) algorithm based on Deep Deterministic Policy Gradient (DDPG); namely Low overhead Multi-Agent DDPG (LoMADDPG) where at each slot, all agents tune their Actor and Critic neural network weights with the help of Global Software-defined Networking (SDN) Controller (GSC) which acts as a global weight tuner. Simulation results illustrate that our proposed task offloading/delivery scheme can result in significant improvement in task completion rate and cost reduction.
11 July 2023 in IEEE Transactions on Green Communications and Networking
With the considerable rise in the number of devices connecting to the Internet, the demand for an extensive capacity communication networks is already apparent. Beam Division Multiple Access (BDMA) is a communication system method aimed at increasing capacity by simultaneously adapting portions of the antenna beam to the locations of multiple Mobile Stations (MSs) achieve enabling more efficient and effective data transmission. However, to assess the best performance, the Base Station (BS) must calculate the exact location of the MSs. In this research, we propose the Direction of Arrival (DoA) approach for locating MSs, where to determine the DoA, an array of antennas is utilized with two common configurations, Uniform Linear Array (ULA) and Uniform Circular Array (UCA). We suggest UCA because the main advantage of the circular array is the 360-degree coverage. Lately, due to the growth of Neural Networks (NNs), many researchers have studied the estimation of DoA using NNs. However, these studies are not widely accepted in practice due to the effect of series errors such as mutual coupling, gain, phase and element position errors and unknown parameters such as the number of input signals and signal level difference. This paper proposes a Deep Neural Network (DNN) framework for estimating the DoA in UCAs. Transfer Learning and Multi-Task techniques eliminate the need to count the number of input signals, reduce errors and bottleneck of NNs make the approach less sensitive than previous techniques. This paper improved the estimation by 75% compared with UCA-MUSIC, 73% compared with the method in (Basikolo et al., 2016), 92% compared with the method in (Hu, 2019), and 90% compared with the method in (Wang et al., 2015) when all the unknown parameters in the system are affected.
28 March 2024 in IEEE Transactions on Network and Service Management
In this paper, fresh Basic Safety Messages (BSM) (e.g., vehicle’s position and speed) are used to control the Connected Automated Vehicles (CAVs) to reduce Time to Collision (TTC) error which leads to decrease in Collision Risk (CR). In contrast to exiting works, a novel Safety-aware Age of Information (S-AoI) metric is proposed that in addition to AoI, takes into account the risk assessment of CAVs to design an efficient transmission protocol for BSMs. We also deploy user-centric Cell-free-massive-MIMO (CFmMIMO) to improve the communication coverage, accessibility, and reliability, where each CAV is served by a cluster of nearby Access Points (APs). Unlike previous works, a two time-scale distributed deterministic policy gradients algorithm is adopted which greatly reduces the signal processing complexity, system load as well as signaling overhead while maintaining the performance. Simulation results show that the proposed framework, i.e, user-centric CFmMIMO technology together with S-AoI metric, can reduce average TTC error between 24%-35% across different lane change probabilities compared to the baseline scenario in which we use small cell mMIMO with AoI metric. Such a reduction in TTC error results in significant decrease (as high as 75%) in CR ratio.
15 October 2024 in IEEE Open Journal of the Communications Society
Network slicing is used to support the stringent requirements of sixth generation (6G) services by dividing an infrastructure network into multiple logical networks that can enable service-oriented resource allocation. However, there are several orchestration issues when considering multiple infrastructure providers (InPs) and multiple tenants in a recursive architecture. There are also challenging issues in designing efficient auction mechanisms for such multi-domain and multi-tenant network slicing. To address these challenges, we consider multi-tenant management and orchestration as a multi-buyer, multi-seller scenario, and propose a novel two-stage auction mechanism that aims to increase the overall utility of all participants while mitigating the overall cost of the network. We formulate this two-stage auction mechanism as a multi-leader multi-follower (MLMF) Stackelberg game approach that converges to a Stackelberg equilibrium. In this game, there are multiple InPs that lease network, computing, and storage infrastructure resources to multiple Tier1 tenants in the first stage of the auction mechanism. Next, Tier1 tenants instantiate triple 6G slices as extremely reliable and low-latency communications (eURLLC), ultra-massive machine-type communications (umMTC), and further enhanced mobile broadband (FeMBB) slices, and lease smaller slices to Tier2 tenants through the second step of the auction mechanism. Tier2 tenants then serve different eURLLC, umMTC, and FeMBB users who have specific and mostly different requirements and constraints, while Tier2 tenants manage their own resources to maximize their utility. Due to the distributed nature of the proposed problem, we consider distributed reinforcement learning (DRL) as a solution. Simulation results show that our DRL-based solution increases the average profit of the network by 19% compared to the existing state-of-the-art benchmark.
5 September 2024 in IEEE Open Journal of the Communications Society
Dynamic spectrum sharing (DSS) is a highly efficient approach for deploying new radio (NR) on lower frequency bands currently utilized by long-term evolution (LTE), to enhance the coverage. This paper addresses several critical issues in DSS. Firstly, we investigate resource block (RB) allocation in a multi-cell environment, considering inter-cell interference (ICI) and frequency reuse (FR) to optimize the combined LTE and NR data rates in a DSS-enabled network. We then apply maximum throughput (MT) and enhanced proportional fairness (ePF) schedulers for RB allocation within our proposed simulation framework. Additionally, we explore the impact of satisfying users quality of service (QoS) on data rate and fairness across various sharing ratio values for LTE and NR guaranteed bit rate (GBR) users. Our results show that while MT achieves higher data rates, ePF ensures better fairness and QoS among users, albeit with a potential data rate reduction of 25-30%. Moreover, under high data rate GBR scenarios, the network can maintain an appropriate fairness index (FI) based on the sharing ratio while guaranteeing GBR users. The ePF scheduler tends to drop more users compared to MT, yet a balance among LTE/NR spectrum sharing ratio, fairness, GBR satisfaction, and overall data rate maximization can be achieved in DSS networks. We also evaluate DSS performance across various realistic propagation models, identifying an optimal sharing ratio that maximizes total data rates for LTE and NR in each environment, with the MT scheduler delivering the highest data rates in rural macro areas. Lastly, we address the issue of demand uncertainty to develop a robust DSS network. Our findings indicate that robust DSS outperforms unrobust DSS by 7-25% and unrobust static spectrum sharing (SSS) by 19-40%.
The dynamic resource block structure (D-RBS) allows for flexible allocation of radio resources. This flexibility (potentially) enables efficient utilization of available resources and adaptability to changing network conditions. In this context, managing resource contention and optimizing allocation decisions become increasingly challenging. In this research paper, we introduce a new approach for D-RBS for re-allocation of enhanced Mobile Broadband (eMBB) and massive Machine Type Communication (mMTC) resource blocs (RBs) to URLLC users. Our scheme leverages artificial intelligence (AI) to support the three main services of 5th generation networks. To efficiently allocate resources for eMBB/mMTC and URLLC services, we propose an intelligent puncturing scheme. Additionally, we formulate an optimization problem that aims to minimize resource and transmit power usage while meeting the requirements of all three services. Since this problem is non-convex and involves multiple optimization variables, we utilize deep reinforcement learning as a solution algorithm. We then compare our proposed intelligent allocation (IA) scheme with two other schemes: random allocation (RaA) and overallocation (OA), which have lower complexity and overhead. Performance and complexity analyses are conducted in a multi-cell scenario with interference. Our results demonstrate that the IA scheme outperforms RaA and OA, achieving an energy efficiency gain of 40% and 15% respectively. However, it is worth noting that IA has a 36.3% higher complexity in terms of action selection compared to RaA and OA.